Incorrect time left and time passed estimate for 1kk iterations with 10kk lines

Problem: Long work with 10kk lines
```
train - 10657047
valid - 4567307
```

```
from catboost import CatBoostClassifier

model = CatBoostClassifier(
iterations=1000000,
early_stopping_rounds=1000,
random_seed=42,
learning_rate=0.6,
custom_loss=['AUC', 'Accuracy', 'F1'],
use_best_model=True,
l2_leaf_reg=3,
bagging_temperature=1,
random_strength=10,
task_type = "GPU"
)
model.fit(
X_train, y_train,
cat_features=cat_features,
eval_set=(X_valid, y_valid),
verbose=False,
plot=True
)
```
54 signs of them 27 categorical signs


catboost version: 0.15.2
Operating System: windows 10
CPU: i7 8700
GPU: 1060 6GB

Start by reproducing the provided CatBoostClassifier configuration with 1,000,000 iterations, the 10-million-line dataset, and GPU execution on the stated environment. Compare the reported time-left and time-passed estimates with the actual run duration; done means the estimates remain accurate for this long-running workload.

Written by the indexing model from the issue text.